Introduction

Carnegie Mellon University (the Contractor) in collaboration with Rensselear Polytechnic Institute, Texas A&M University, University of Minnesota, and University of Illinois at Urbana-Champaign shall establish a consortium involving researchers from the five universities. The purpose of this effort is to develop a new type of agent, the context dependent network (CDN) agent, whose deployment will improve the agility and robustness (survivability) of large-scale dynamic networks that face new and unanticipated operating conditions. The target network will be the U.S. power grid, one of the critical components of our national infrastructure. Some of the concepts and tools developed will be specific to this network, but many will be general to the class to which the power grid belongs, namely, networks whose operation involves large numbers of distributed control devices and whose sensors produce a continual stream of large sets of data. Traffic networks and telecommunication networks are other examples from this class of networks.

Background

The need for context-dependent agents is clear: the networks considered here are neither uniform in space nor time. Rather, their structural properties vary with location and change when equipment is added or removed. Moreover, their states can vary profoundly in both space and time: at any instant, some parts of a large network may be operating normally while others are profoundly abnormal and rapidly expanding. In other words, the ill effects of a contingency (a natural disturbance or a deliberate attack) can spread rapidly. Countering these effects requires quick diagnosis and control. In the past, this quick reaction has been achieved by preprogramming the responses of the front-line of the networks defense (distributed monitoring and control devices such as relays and voltage regulators). Adaptive responses have been confined to centralized facilities, such as energy management centers, that, by their centralized nature, are slow. Therefore, existing networks are unable to provide both fast and suitably adaptive (context- dependent) responses to contingencies.

In power networks, the pre-programmed responses of the front-line devices to contingencies have, in almost all cases, proved adequate for avoiding serious consequences. However, on occasions that in the past were rare but now seem to be occurring with greater frequency, the contingencies trigger a sequence of cascading failures. With the network in disarray, pre-programmed control actions tend to be ineffective and can even cause further problems. System operators watch helplessly as the system breaks up into several pieces, leaving many customers without service. A post-mortem analysis invariably indicates that one of several control actions, if it had been "in place," would have stopped the network from breaking up. Usually, the flows on the network are restricted, burdening consumers with higher costs, until one of these control actions can be put "in place" to ensure that the sequence of events that caused the break-up will not be repeated. (The reduced power transfer in the WSCC system after the breakup incidents in 1996 has been estimated to cost the customers between 2-5 billion dollars.) However, as the loading on the network increases, these piece-meal countermeasures will no longer offer the same security margin.  It is only a matter of time before more and bigger break-ups occur.

We believe that one great drawback to the ISO (independent system operator) methodology, now being implemented in the US, is the degree to which it further centralizes transmission control and thereby, decreases system agility. To make systems much more agile and to "leapfrog" the traditional approach to making control improvements, we propose a relatively inexpensive solution: the deployment of CDN agents throughout the network. These agents will be able to act autonomously, and therefore, quickly. They will be able to learn from their real and simulated experiences, and therefore, each will continually improve its competence to deal with its context. In other words, with only a relatively small investment in computing and communication equipment, a CDN agent can be added to any network component, giving the component the advantages of decision-making that is autonomous when necessary, context dependent and self-improving.

We envision the massive deployment of CDN agents in power networks, so centralized facilities, such as ISO, can be relieved of all but the highest level of decision-making. The CDN agents would implement these high-level decisions in the contexts of the network-components to which they are assigned. In addition, each agent would diagnose the condition of its "component," call for assistance when necessary, diagnose the mode (condition) of the network in its neighborhood, and react quickly and appropriately to this mode. The agent would spend the time between these jobs automatically learning how to do them better, using actual or simulated network data (it could perform the simulations itself or request them from a centralized simulation-facility).

Of course, these CDN agents will vary by network-component. For instance, the agent for a relay will be simpler than the agent for a generator. Therefore, any massive deployment is conditional on a streamlined process for assembling a variety of CDN agents. Other obstacles to realizing our vision are: the new operating modes of restructured power systems are not
well understood; the numbers of variables in power system state descriptions are too large for the direct application of existing automatic learning algorithms; distributed agents, when they act autonomously, must have some way of coordinating their actions so they do not work at cross-purposes but towards global goals (which they might not explicitly know); the agents will require a robust, real-time operating system; no single university has the skills to overcome the aforementioned obstacles; and multi-university consortia have, in the past, been unable to develop the level of collaboration that can occur among researchers that are close together. Our approach to overcoming these obstacles and carrying our vision to the point of demonstrating CDN agents in representative power networks, consists of six
parallel tasks.

Techniques

Modeling
        operating modes
        contingencies
        impact of restructured power systems
        device capabilities/influence

State estimation
        using local information
        network state estimation
        real-time constraints

Hybrid control
        adaptive mode switching
        coverage

Learning
        distributed learning
        state-space decomposition

Coordination
        collaboration strategies
        moving off-line techniques for asynchronous algorithms on-line

Objectives